ProBASS-a language model with sequence and structural features for predicting the effect of mutations on binding...

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Mutations in PPIs can lead to diseases.
• Existing methods for predicting binding affinity changes (ΔΔGbind) often lack precision.
• Protein language models (PLMs) show promise but require optimization for ΔΔGbind prediction.

Purpose of the Study:

• To develop a precise and broadly applicable model for predicting mutation effects on protein binding affinity (ΔΔGbind).
• To leverage advanced protein language models (PLMs) for improved ΔΔGbind prediction accuracy.

Main Methods:

• Developed ProBASS (Protein Binding Affinity from Structure and Sequence) approach.
• Utilized two advanced PLMs (ESM2, ESM-IF1) for sequence and structural data.
• Generated embeddings for PPI mutants and fine-tuned the model on experimental ΔΔGbind data.

Main Results:

• ProBASS achieved high correlations with experimental ΔΔGbind values (0.83±0.05 for single, 0.69±0.04 for double mutations on same PPI).
• On a large dataset of 2,325 single mutations, ProBASS reached a correlation of 0.81±0.02.
• Significantly outperformed other PLMs in predictive accuracy for ΔΔGbind.

Conclusions:

• Refining pre-trained PLMs with extensive ΔΔGbind datasets improves prediction accuracy.
• ProBASS offers a precise and broadly applicable tool for predicting mutation effects on binding affinity.
• The model facilitates future protein engineering and design studies and can be further improved with more data.